| Term | Definition |
|---|---|
| Partner API key | A credential shaped sp_<key_id>_<secret> that identifies an organization, not a person. Your backend sends it as a bearer token. |
| End user | One of your own customers, named on a request with X-SP-End-User. Priors hashes the value and does not store it. |
| Organization | Your company. Your key resolves to it, and it owns every agent the key creates. |
| Agent visibility | Every agent your key creates is external: owned by your organization, in no directory, reachable only by a key of that organization. |
| Agent | One memory with a role and an objective. An account creates it and owns it. |
| Thread | One conversation with an agent. It gives recent context. It belongs to the subject that created it: your organization, or the person you named with X-SP-End-User. |
| Message | One text in a thread, from the user or from the agent. The API sometimes calls this a turn. |
| Speaker | Who said a turn: the speaker field on an observation. Without it, the first entry of tags (older behavior), and user when there are no tags either. |
| Tags | Free-form labels on an observation, such as the channel or topic. Kept with the turn when speaker is sent. |
| Detail | How much an answering call writes: brief, standard (the default), or detailed. An instruction to the agent, never a cut. |
| Not learned | A turn whose memory could not be written. It is never learned later. Responses report it in learning, with any refund in refunded_credits. Send it again. |
| Observation | One thing for the agent to learn from: a list of content blocks, with an optional time, labels, and file ids. |
| Content block | One part of an observation. Today always {"type": "text", "text": "..."}. |
| Observation request | One observe or observe-file call. It has an observation_id and a status. |
| Poll | To call GET on a status again and again, for example every second, until the status changes. |
| Committed | The status value that means every memory from the observation request has been written. |
| Memory | One fact that the agent wrote from an observation. It has the text of the fact and the time the fact refers to. You cannot read memories directly. |
| Memory id | The mem_... id of a memory. Answers return these ids, so you can see which memories were used. |
| Memory search | The step in every answering call that finds the memories that relate to the message. |
| Context | The memory text that observe-and-enrich returns for your own model. |
| Answer | The reply that observe-and-respond or chat writes. |
| Source | A file that the agent holds: a document it learned from, or an image attached to a chat message. |
| Citation | A link from an answer to the source it used. |
| Piece | One part of a split document. turns_total counts the pieces. |
| Text recognition | Reading the text in a scanned page or an image. Also called OCR. |
| Idempotency key | A value you send with an upload so that sending the same upload again does not read or charge the file a second time. |
| Duplicate | A file whose bytes this agent already read. In the same thread it returns 202 with duplicate: true and costs nothing. In another thread it returns 409 document_already_ingested unless you send allow_duplicate=true. |
| Rate limit | The largest number of requests allowed in a time window. |
| Credits | The unit in which Priors charges for use. |
| Request id | The request_id in every error. Send it to support when you report a problem. |
Reference
Terminology
Every word this documentation uses, defined once.
This page defines every term that the other pages use. Each term has one meaning, and the other pages use the same word for the same thing.